Falsification-First AI for Biological Target Assessment: Lessons from Four Prospectively Frozen Experiments

AI systems can produce plausible biological stories faster than people can test them. This report asks a narrower question: can a research workflow make unsupported conclusions harder to keep? We report four prospectively bounded experiments. In Pilot 2, a strong agentic baseline recovered known biology better than a heavier architecture while matching it on four frozen discipline dimensions. A later Lean Scientific Agent Core preserved specified safeguards in one fresh-domain transfer. In H3-v2, an ulcerative-colitis causal-chain study, the focal reading weakened under two adversarial checks and the programme ended as not evaluable on available public data. In a separate TTPAL challenge, one gut-resident T-cell discovery resource supported TTPAL in 3 of 66 designated groups, with PP4 up to about 0.992. PP4 is a model score for whether two association patterns look consistent with sharing a signal. It is not the chance that TTPAL causes disease. The blood data had no usable TTPAL expression quantitative trait locus (eQTL) signal; across the six measured blood T-cell classes in the primary independent cohort, PP4 was at most 0.031. The blood result does not settle whether a gut-resident effect exists. These results do not validate a therapeutic target, clinical intervention, diagnostic, safety claim or general-purpose drug-discovery system. They show how frozen questions, lineage accounting and stopping rules can make negative, discordant and not-evaluable outcomes useful.

Version 1.1 of a falsification-first report of four prospectively frozen computational experiments. The typeset PDF is the paper; the Markdown file is the same manuscript as open plain text; the reproducibility supplement holds the manuscript source, evidence substrate, LaTeX build and quality-assurance programs, and verifies itself after extraction. Version 1.1 revises typesetting, section structure, figures, in-text citations and packaging and adds competing-interest, ethics and data-availability declarations, and every changed sentence is listed in the supplement's editorial change register; the scientific content, results and claim boundaries of version 1.0 are unchanged. The report preserves negative, discordant and not-evaluable outcomes and does not validate a therapeutic target, clinical intervention, diagnostic, safety claim or general-purpose drug-discovery system.